Abstract: To address the inadequate characterization of extreme operating scenarios under high-penetration wind and photovoltaic integration and the potential failure of conventional load-flow algorithms under critical operating conditions, this paper proposes a comprehensive renewable energy hosting capacity evaluation method based on the holomorphic embedding load-flow method (HELM). An adversarial scenario generative adversarial network (ASGAN) is employed to learn the spatio-temporal correlations, statistical characteristics, and physical limits of wind and photovoltaic power outputs and to generate candidate scenarios. The candidate scenarios are then evaluated using the holomorphic embedding load-flow model. A comprehensive risk index is constructed from bus-voltage violations, branch overloads, and power-supply deficits, and the resulting risk information is fed back to the generator through a surrogate risk evaluator, thereby forming a closed-loop training mechanism integrating scenario generation and load-flow risk assessment. A high-risk joint-scenario screening and probability-correction method is further developed, and the maximum renewable energy hosting capacity is determined using the loss-of-load probability (LOLP) and expected energy not supplied (EENS). Finally, case studies on the IEEE 118-bus system demonstrate that under the LOLP limit of 0.1%, compared with the conventional method combining generative adversarial network and HELM, the proposed approach can avoid overestimation of renewable energy hosting capacity caused by the omission of low-probability high-risk scenarios and improve the reliability of renewable energy hosting capacity evaluation.